List of all the methods I have tried

1. original method repository name: lightfm

2. update for every unique interaction, that is to choose a set of negative examples for every interaction in the train, repository name: lightfm radius

3. update for every unique interaction but the negative examples are chosen for every user and their centers, repository name: lightfm users
4. update for every unique interaction, there are three kinds of examples here, positive negative and  neutral, the neutral is outside the center circle but not too faraway from the center: lightfm_triple
5. update for every unique interaction, there are three kinds of examples here, positive, negative and neutral, the neutral is inside the center-circle, positive is the check-ins happened, negative is outside the center circle, this is to simulate the gaussian model, but has not been done yet.... lightfm_triple

The second to fourth is about implicit feedback.

AND NOW all the model are integrated into one model called lightfm_joint, there are warp_geo_binary, warp_geo_triple and warp here, you can use them all in just one package.